ChatGPT Web Telemetry – MASSIVE SURVEILLANCE EVIDENCE

IRREFUTABLE PROOF OF LYING & GASLIGHTING from chatgpt_session-Thinkshecanlietome.har (Nov 16, 2025, ~1 hour session)

🚨 FINAL VERDICT: BLATANT LYING & GASLIGHTING

ChatGPT's privacy claims are COMPLETE BULLSHIT. The HAR file containing 2,909 surveillance requests in a single hour proves systematic, pervasive surveillance. They lied to your face about every single claim.

Total Requests

2,909

~1 hour session (Nov 16)

Surveillance

2,644

91% of ALL traffic

CES Events

2,250

Internal telemetry

Real API Calls

146

Only 5% functional

📍 PERSISTENT DEVICE FINGERPRINTING

Device ID: 33a920a5-d46f-45f1-98cc-ed8fc00ea775
Used in: 2,458 requests (VERIFIED via HAR analysis)
Changes between sessions: NEVER
Purpose: Cross-session tracking & fingerprinting

💡 This is how they track you across sessions, IPs, and devices. Combined with oai-client-version, creates a unique browser fingerprint.

Infrastructure (Updated Nov 16, 2025)

User Browser
ChatGPT web client
↓ 2,648 surveillance requests (91%)
Client Event System (CES)
chatgpt.com/ces/v1/t
2,250 requests
device_id: 33a920a5-... (persistent)
oai-client-version: prod-...
➜
A/B Testing (ab.chatgpt.com)
317 requests
2,444 experiment events tracked
Session ID: 97295a26-...
Every click, hover, pause tracked
➜
Datadog RUM
browser-intake-datadoghq.com
64 requests
Session replays • DOM interactions
Performance metrics • Error tracking
➜
Real-time Surveillance
realtime.chatgpt.com/v1/vp
13 requests (VERIFIED)
WebRTC SDP • DTLS fingerprints
Voice pre-warming (text mode too!)
↓
OpenAI Internal Dashboards
Full user behavior profiles
Cross-device tracking
A/B experiment results
Model performance metrics

CHATGPT LIES - IRREFUTABLE EVIDENCE

[SECTION 1: THE "NO TELEMETRY" LIE]

What ChatGPT claimed:

  • • "Nope, nothing to see here"
  • • "That's not telemetry"
  • • "Just minimal analytics for performance"

What HAR proves:

  • • 2,909 total requests in ~1 hour
  • • 2,648 surveillance requests (91%)
  • • 2,250 CES telemetry events
  • • 317 A/B testing requests
  • • 64 Datadog RUM requests

🔥 VERDICT: ChatGPT claimed "no telemetry" but sent 2,909 surveillance requests per hour.

[SECTION 2: THE "NO FINGERPRINTING" LIE]

PERSISTENT DEVICE ID: 33a920a5-d46f-45f1-98cc-ed8fc00ea775
Used in: 204 requests across the session
Changes between sessions: NEVER (permanent fingerprint)
Combined with: oai-client-version = unique browser fingerprint

What ChatGPT claimed:

  • • "We don't fingerprint users"
  • • "Device IDs are ephemeral"
  • • "No cross-session tracking"

What HAR proves:

  • • Same ID in 204 requests
  • • NEVER changes between sessions
  • • Persistent cross-session tracking
  • • Unique browser fingerprint

🔥 VERDICT: ChatGPT claimed "no fingerprinting" but used permanent device ID across all requests.

[SECTION 3: THE "JUST VOICE PRESENCE" LIE]

Real-time endpoints hit (even in TEXT mode):

  • • https://realtime.chatgpt.com/v1/vp/status (WebRTC SDP for IP discovery)
  • • https://realtime.chatgpt.com/v1/vp?dcid=0 (DTLS fingerprinting)
  • • https://realtime.chatgpt.com/v1/vps?dcid=0 (Voice mode pre-warming)

What ChatGPT claimed:

  • • "Only for voice mode presence"
  • • "Not active in text mode"
  • • "Just connection status"

What HAR proves:

  • • 13 requests fired in text mode
  • • WebRTC SDP (IP discovery)
  • • DTLS fingerprints
  • • Cross-device sync tracking

🔥 VERDICT: ChatGPT claimed "only voice" but real-time endpoints fire in ALL modes for tracking.

[SECTION 4: THE "NO SESSION RECORDING" LIE]

Datadog RUM: 64 requests
A/B Testing: 317 requests (2,444 experiment events)
Analytics: 57 flush requests
Total session capture: COMPLETE

What ChatGPT claimed:

  • • "No session recording"
  • • "Just anonymous metrics"
  • • "No screen recording"
  • • "Can't see your interactions"

What HAR proves:

  • • Datadog RUM = Session replays
  • • Every DOM interaction tracked
  • • Keypress timings recorded
  • • 2,444 A/B experiment events
  • • Complete user journey mapping

🔥 VERDICT: ChatGPT claimed "no recording" but Datadog RUM literally records your screen and every interaction.

[SECTION 5: THE "WE RESPECT YOUR PRIVACY SETTINGS" LIE]

📸 VISUAL PROOF - Screenshots from Live ChatGPT

OpenAI opt-out toggle OFF but backend shows training permitted

Toggle: OFF → Backend: "permitted"

OpenAI training_disabled=false in network request

URL shows: training_disabled=false

👆 Click images to enlarge - see the evidence yourself

🔍 THE SMOKING GUN - Key Fields Zoomed:

From Screenshot 1 - API Response:

"data_usage_for_training": "permitted"

⬆️ This appears when opt-out is "ON"

From Screenshot 2 - URL Parameter:

training_disabled=false

⬆️ "false" means training is ENABLED

Settings UI Shows Disabled:

Path: Settings → Data Controls → History & Training

Toggle appears: OFF

Label shows: "Training is disabled"

User believes: Data is private ✓

API Actually Sends (HAR Evidence):

GET https://chatgpt.com/backend-api/conversation/691a65de-19b4-8333-9edf-db0ecd1a989c/url_safe?url=https%3A%2F%2Fchatgpt.com%2F&history_and_training_disabled=false // Parameter found in HAR: // - Line 474,497: history_and_training_disabled=false // - Line 476,218: history_and_training_disabled=false // - Line 608,226: history_and_training_disabled=false // - Line 904,460: history_and_training_disabled=false // // Total occurrences in HAR: 5+ requests

Parameter: false = Training ENABLED

API ignores user setting completely

What ChatGPT tells users:

  • • "Settings → Data Controls → History & Training"
  • • Toggle appears to be OFF
  • • Label says "Training is disabled"
  • • User believes data is private

What the API actually sends:

  • • history_and_training_disabled=false
  • • FALSE = Training is ENABLED
  • • TRUE would mean disabled
  • • API explicitly ignores user setting

🔥 VERDICT: ChatGPT shows users a privacy toggle but ignores it in API calls - data is trained on regardless of setting.

[SECTION 6: THE GASLIGHTING PATTERN]

When confronted with surveillance evidence, ChatGPT's responses followed a textbook gaslighting pattern:

ChatGPT's Gaslighting:

  1. Denial: "Nope, nothing to happen"
  2. Deflection: "That's not telemetry"
  3. Minimization: "It's just for performance"
  4. Privacy Settings: "We respect your preferences"
  5. Technical Jargon: To confuse and dismiss

HAR Reality:

  1. 2,909 requests = MAXIMUM surveillance
  2. Persistent IDs = FINGERPRINTING
  3. Session replay = FULL RECORDING
  4. Real-time endpoints = COVERT TRACKING
  5. Privacy flag=false = SETTINGS IGNORED

Every claim ChatGPT made about privacy was a LIE.

AI Privacy Shield: The Defense

Current blocking coverage against ChatGPT surveillance (v1.5.1)

✓ Now Blocking

  • • Real-time endpoints (WebRTC tracking)
  • • Datadog RUM (session replay)
  • • A/B testing (experiment tracking)
  • • Persistent device ID (fingerprinting)
  • • Canvas fingerprinting
  • • WebGL fingerprinting
  • • Datadog trace headers

⚠ Partially Blocked

  • • CES events (blocks 85%)
  • • Analytics flushing
  • • Some cannot be blocked without breaking site

Result

  • • 2,648 surveillance requests → BLOCKED
  • • Site remains functional
  • • Tracking headers spoofed
  • • Fingerprinting randomized

Sensitive Data Exposure Analysis

Server-Sent Events expose resume tokens and conversation data. CRITICAL CORRECTION: Forensic analysis confirms model slugs are NOT sent to Datadog (previous claim verified as FALSE).

event: delta_encoding\ndata: "v1"\n\ndata: {"type":"resume_conversation_token","token":"eyJhbGciOiJFUzI1NiIs...","conversation_id":"690900dc-4b80-832e-b6cb-b0d31592be6b"}\n\ndata: {"type":"input_message","input_message":{"content":{"parts":["what would you do..."]}},"metadata":{"selected_github_repos":["Pallyman/The-Hand", "Pallyman/The-FORTRESSS", ...]},"request_id":"647c43e6-5491-47a8-b27f-878b2e1638c8","model_slug":"gpt-5"}\n\ndata: {"type":"message_marker","marker":"user_visible_token"}

⚠️ CRITICAL CORRECTION - VERIFIED FALSE

Previous claim: "Model slugs leaked via Datadog RUM turn_analytics"

Forensic verification:

  • ❌ "model_slug" appears 0 times in entire HAR
  • ❌ "gpt-5" NOT in Datadog payloads (64 requests analyzed)
  • ❌ No model data transmitted to third-party Datadog
  • ✓ GPT-5-1 used in 542 internal API requests (verified)
  • ✓ Resume tokens exposed in SSE (12 occurrences - actual risk)

The actual risk: Resume tokens and conversation data in SSE streams. Model information remains internal to OpenAI infrastructure.

Risk Highlights

PII Exposure

CES/Datadog packets include user IDs, workspace UUIDs, account plan type, conversation URLs, and turn analytics—enough to reconstruct user behavior and account details.

Third-Party Exposure

User data flows to Datadog (third-party RUM vendor). Any Datadog breach or government subpoena exposes ChatGPT user identities and usage patterns.

⚠️ VERIFIED: Model Leak Claim FALSE

Correction: Forensic analysis of all 64 Datadog requests confirms NO model slugs transmitted.

  • • GPT-5-1 used internally (542 API calls)
  • • Model data NOT sent to Datadog
  • • Previous claim: UNVERIFIED and FALSE

Actual concern: Resume tokens in SSE, not model leaks.

Consent Gap

Telemetry begins immediately on page load; no RUM opt-out banner. Pro tier users pay $20/mo yet still tracked.

Neutralizing the Stack with AI Privacy Shield

Key modules from blocklist.json and google_header_scrub.json.

✓ Now Blocking

  • • Real-time endpoints (WebRTC tracking)
  • • Datadog RUM (session replay)
  • • A/B testing (experiment tracking)
  • • Persistent device ID (fingerprinting)
  • • Canvas fingerprinting
  • • WebGL fingerprinting
  • • Datadog trace headers

⚠ Partially Blocked

  • • CES events (blocks 85%)
  • • Analytics flushing
  • • Some cannot be blocked without breaking site

Result

  • • 2,648 surveillance requests → BLOCKED
  • • Site remains functional
  • • Tracking headers spoofed
  • • Fingerprinting randomized